{
  "id": 126534,
  "title": "[Solved] My Every Submission got score 0.0614 ...",
  "url": "/competitions/bengaliai-cv19/discussion/126534",
  "author_name": "kaerururu",
  "post_date": "2020-01-18T06:35:10.934000",
  "votes": 2,
  "comment_count": 19,
  "views": 0,
  "content": "<p>I try to make baseline resnet18 model with 100 epoch training, Adam(4e-4), CELoss, no Augmentations.</p>\n\n<p>But every experiments result in sumple_submission score 0.0614 ..</p>\n\n<p>What happens with me !!! Thank you !</p>",
  "messages": [
    {
      "id": 722351,
      "postDate": "2020-01-18T12:51:28.053Z",
      "content": "<p>Something like that can also happen if you forget to normalize your images in your code. As model is trained on normalized data. So check that also.</p>",
      "rawMarkdown": "Something like that can also happen if you forget to normalize your images in your code. As model is trained on normalized data. So check that also.",
      "votes": 1,
      "replies": [
        {
          "id": 722554,
          "postDate": "2020-01-18T17:49:14.507Z",
          "content": "<p>```\nclass BengaliAIDatasetTest(torch.utils.data.Dataset):\n    def <strong>init</strong>(self,df,transform=None):\n        self.df = df\n        self.transform = transform</p>\n\n<pre><code>def __len__(self):\n    return len(self.df)\n\ndef __getitem__(self,idx):\n\n    input_dic = {}\n    image = self.df.iloc[idx][1:].values.reshape(128,128).astype(np.float)\n    image = threshold_image(image)\n    image = self.transform(image=image)['image']\n    image = (image.astype(np.float32) - 0.0692) / 0.2051\n    image = image_to_tensor(image, normalize=False) \n\n    input_dic['image'] = image\n\n    return input_dic\n</code></pre>\n\n<p>```</p>\n\n<p>In my TestDataset, I normalize images below code</p>\n\n<p><code>\n image = (image.astype(np.float32) - 0.0692) / 0.2051\n</code> </p>\n\n<p>Does this work unlike my expectation?</p>",
          "rawMarkdown": "```\nclass BengaliAIDatasetTest(torch.utils.data.Dataset):\n    def __init__(self,df,transform=None):\n        self.df = df\n        self.transform = transform\n        \n    def __len__(self):\n        return len(self.df)\n    \n    def __getitem__(self,idx):\n        \n        input_dic = {}\n        image = self.df.iloc[idx][1:].values.reshape(128,128).astype(np.float)\n        image = threshold_image(image)\n        image = self.transform(image=image)['image']\n        image = (image.astype(np.float32) - 0.0692) / 0.2051\n        image = image_to_tensor(image, normalize=False) \n        \n        input_dic['image'] = image\n        \n        return input_dic\n```\n\nIn my TestDataset, I normalize images below code\n\n```\n image = (image.astype(np.float32) - 0.0692) / 0.2051\n``` \n\nDoes this work unlike my expectation?"
        }
      ]
    },
    {
      "id": 722238,
      "postDate": "2020-01-18T09:51:38.923Z",
      "content": "<p>It happened to mee too. In your submission kernel, try to use the train parquet files for a test commit. </p>",
      "rawMarkdown": "It happened to mee too. In your submission kernel, try to use the train parquet files for a test commit. ",
      "votes": 1,
      "replies": [
        {
          "id": 722553,
          "postDate": "2020-01-18T17:45:07.207Z",
          "content": "<p>Thank you <a href=\"/pestipeti\">@pestipeti</a> !</p>\n\n<p>I'll try using original parquet files in next experiment.</p>",
          "rawMarkdown": "Thank you @pestipeti !\n\nI'll try using original parquet files in next experiment."
        },
        {
          "id": 722992,
          "postDate": "2020-01-19T11:23:26.820Z",
          "content": "<p>Using original parquet files brought me 'reasonable' score!</p>\n\n<p>Your advice helps me a lot, thx. m(_ _)m</p>",
          "rawMarkdown": "Using original parquet files brought me 'reasonable' score!\n\nYour advice helps me a lot, thx. m(_ _)m",
          "votes": 1
        }
      ]
    },
    {
      "id": 722140,
      "postDate": "2020-01-18T06:55:10.937Z",
      "content": "<p>Maybe there is some problem with your submission kernel??</p>",
      "rawMarkdown": "Maybe there is some problem with your submission kernel??",
      "votes": 1
    },
    {
      "id": 722375,
      "postDate": "2020-01-18T13:47:13.867Z",
      "content": "<p>During submission the test_images perquet files will be replaced by the complete one. So make your submission for the full test data</p>",
      "rawMarkdown": "During submission the test_images perquet files will be replaced by the complete one. So make your submission for the full test data",
      "votes": 2
    },
    {
      "id": 722135,
      "postDate": "2020-01-18T06:44:18.197Z",
      "content": "<p>Hi,</p>\n\n<p>Are you deriving test values from test dataset or from sample_submission file, if you predicting values from sample_submission than replace it with data from test parquet files.</p>\n\n<p>Also it might happen when the algorithm is overfitted than it gives score of .0614.</p>",
      "rawMarkdown": "Hi,\n\nAre you deriving test values from test dataset or from sample_submission file, if you predicting values from sample_submission than replace it with data from test parquet files.\n\nAlso it might happen when the algorithm is overfitted than it gives score of .0614.",
      "votes": 2,
      "replies": [
        {
          "id": 722248,
          "postDate": "2020-01-18T10:05:15.277Z",
          "content": "<p>Oh yeah. Thank you so much. I just realized that sample submission file will contain only 11 rows, whereas test part that is hidden from us is definitely much bigger.</p>",
          "rawMarkdown": "Oh yeah. Thank you so much. I just realized that sample submission file will contain only 11 rows, whereas test part that is hidden from us is definitely much bigger."
        },
        {
          "id": 722305,
          "postDate": "2020-01-18T11:24:33.003Z",
          "content": "<p>That's great <a href=\"/nroman\">@nroman</a> </p>",
          "rawMarkdown": "That's great @nroman "
        },
        {
          "id": 722555,
          "postDate": "2020-01-18T17:52:28.507Z",
          "content": "<p>Overfitting model gives me 0.0614 ??</p>",
          "rawMarkdown": "Overfitting model gives me 0.0614 ??",
          "votes": 1
        },
        {
          "id": 722560,
          "postDate": "2020-01-18T18:03:45.367Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    },
    {
      "id": 722132,
      "postDate": "2020-01-18T06:35:10.933Z",
      "content": "<p>I try to make baseline resnet18 model with 100 epoch training, Adam(4e-4), CELoss, no Augmentations.</p>\n\n<p>But every experiments result in sumple_submission score 0.0614 ..</p>\n\n<p>What happens with me !!! Thank you !</p>",
      "rawMarkdown": "I try to make baseline resnet18 model with 100 epoch training, Adam(4e-4), CELoss, no Augmentations.\n\nBut every experiments result in sumple_submission score 0.0614 ..\n\nWhat happens with me !!! Thank you !\n\n",
      "votes": 2
    },
    {
      "id": 722997,
      "postDate": "2020-01-19T11:29:18.303Z",
      "content": "<p>Thank you for a lot of advices.</p>\n\n<p>In my case, Not using origianl test parquet but using transformed feather files resluted in this bag. </p>",
      "rawMarkdown": "Thank you for a lot of advices.\n\nIn my case, Not using origianl test parquet but using transformed feather files resluted in this bag. "
    },
    {
      "id": 722552,
      "postDate": "2020-01-18T17:42:54.537Z",
      "content": "<p>Thank you for many comments !!\nBelow, my latest submission code is here.</p>\n\n<p>some advices may help me... thank you.</p>\n\n<p><a href=\"https://www.kaggle.com/kaerunantoka/pred-bengali-resnet18-exp5?scriptVersionId=27177286\">https://www.kaggle.com/kaerunantoka/pred-bengali-resnet18-exp5?scriptVersionId=27177286</a></p>",
      "rawMarkdown": "Thank you for many comments !!\nBelow, my latest submission code is here.\n\nsome advices may help me... thank you.\n\nhttps://www.kaggle.com/kaerunantoka/pred-bengali-resnet18-exp5?scriptVersionId=27177286",
      "replies": [
        {
          "id": 722574,
          "postDate": "2020-01-18T18:35:38.427Z",
          "content": "<p>You are using sample submission file. Use test set, combined from 4 parquet files instead.</p>",
          "rawMarkdown": "You are using sample submission file. Use test set, combined from 4 parquet files instead.",
          "votes": 2
        },
        {
          "id": 722711,
          "postDate": "2020-01-19T00:45:59.023Z",
          "content": "<p>Oh.. I see. I'll try it, thx!!</p>",
          "rawMarkdown": "Oh.. I see. I'll try it, thx!!"
        },
        {
          "id": 722991,
          "postDate": "2020-01-19T11:21:55.420Z",
          "content": "<p>Thank you for your advice!!</p>\n\n<p>Using not feather files but test parquet files resulted in 'reasonable' score. m(_ _)m</p>",
          "rawMarkdown": "Thank you for your advice!!\n\nUsing not feather files but test parquet files resulted in 'reasonable' score. m(_ _)m"
        }
      ]
    },
    {
      "id": 722245,
      "postDate": "2020-01-18T10:00:14.110Z",
      "content": "<p>Same for me!</p>",
      "rawMarkdown": "Same for me!"
    },
    {
      "id": 722260,
      "postDate": "2020-01-18T10:24:00.340Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 722351,
      "author_name": "Gaurav Gautam",
      "author_url": "",
      "post_date": "2020-01-18T12:51:28.053000",
      "content": "<p>Something like that can also happen if you forget to normalize your images in your code. As model is trained on normalized data. So check that also.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 722554,
          "author_name": "kaerururu",
          "author_url": "",
          "post_date": "2020-01-18T17:49:14.507000",
          "content": "<p>```\nclass BengaliAIDatasetTest(torch.utils.data.Dataset):\n    def <strong>init</strong>(self,df,transform=None):\n        self.df = df\n        self.transform = transform</p>\n\n<pre><code>def __len__(self):\n    return len(self.df)\n\ndef __getitem__(self,idx):\n\n    input_dic = {}\n    image = self.df.iloc[idx][1:].values.reshape(128,128).astype(np.float)\n    image = threshold_image(image)\n    image = self.transform(image=image)['image']\n    image = (image.astype(np.float32) - 0.0692) / 0.2051\n    image = image_to_tensor(image, normalize=False) \n\n    input_dic['image'] = image\n\n    return input_dic\n</code></pre>\n\n<p>```</p>\n\n<p>In my TestDataset, I normalize images below code</p>\n\n<p><code>\n image = (image.astype(np.float32) - 0.0692) / 0.2051\n</code> </p>\n\n<p>Does this work unlike my expectation?</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 722238,
      "author_name": "Peter",
      "author_url": "",
      "post_date": "2020-01-18T09:51:38.923000",
      "content": "<p>It happened to mee too. In your submission kernel, try to use the train parquet files for a test commit. </p>",
      "votes": 1,
      "replies": [
        {
          "id": 722553,
          "author_name": "kaerururu",
          "author_url": "",
          "post_date": "2020-01-18T17:45:07.207000",
          "content": "<p>Thank you <a href=\"/pestipeti\">@pestipeti</a> !</p>\n\n<p>I'll try using original parquet files in next experiment.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 722992,
          "author_name": "kaerururu",
          "author_url": "",
          "post_date": "2020-01-19T11:23:26.820000",
          "content": "<p>Using original parquet files brought me 'reasonable' score!</p>\n\n<p>Your advice helps me a lot, thx. m(_ _)m</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 722140,
      "author_name": "Bibek",
      "author_url": "",
      "post_date": "2020-01-18T06:55:10.937000",
      "content": "<p>Maybe there is some problem with your submission kernel??</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 722375,
      "author_name": "Bibhas Mondal",
      "author_url": "",
      "post_date": "2020-01-18T13:47:13.867000",
      "content": "<p>During submission the test_images perquet files will be replaced by the complete one. So make your submission for the full test data</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 722135,
      "author_name": "Amit",
      "author_url": "",
      "post_date": "2020-01-18T06:44:18.197000",
      "content": "<p>Hi,</p>\n\n<p>Are you deriving test values from test dataset or from sample_submission file, if you predicting values from sample_submission than replace it with data from test parquet files.</p>\n\n<p>Also it might happen when the algorithm is overfitted than it gives score of .0614.</p>",
      "votes": 2,
      "replies": [
        {
          "id": 722248,
          "author_name": "Roman",
          "author_url": "",
          "post_date": "2020-01-18T10:05:15.277000",
          "content": "<p>Oh yeah. Thank you so much. I just realized that sample submission file will contain only 11 rows, whereas test part that is hidden from us is definitely much bigger.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 722305,
          "author_name": "Amit",
          "author_url": "",
          "post_date": "2020-01-18T11:24:33.003000",
          "content": "<p>That's great <a href=\"/nroman\">@nroman</a> </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 722555,
          "author_name": "kaerururu",
          "author_url": "",
          "post_date": "2020-01-18T17:52:28.507000",
          "content": "<p>Overfitting model gives me 0.0614 ??</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 722560,
          "author_name": "",
          "author_url": "",
          "post_date": "2020-01-18T18:03:45.367000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 722997,
      "author_name": "kaerururu",
      "author_url": "",
      "post_date": "2020-01-19T11:29:18.303000",
      "content": "<p>Thank you for a lot of advices.</p>\n\n<p>In my case, Not using origianl test parquet but using transformed feather files resluted in this bag. </p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 722552,
      "author_name": "kaerururu",
      "author_url": "",
      "post_date": "2020-01-18T17:42:54.537000",
      "content": "<p>Thank you for many comments !!\nBelow, my latest submission code is here.</p>\n\n<p>some advices may help me... thank you.</p>\n\n<p><a href=\"https://www.kaggle.com/kaerunantoka/pred-bengali-resnet18-exp5?scriptVersionId=27177286\">https://www.kaggle.com/kaerunantoka/pred-bengali-resnet18-exp5?scriptVersionId=27177286</a></p>",
      "votes": 0,
      "replies": [
        {
          "id": 722574,
          "author_name": "Roman",
          "author_url": "",
          "post_date": "2020-01-18T18:35:38.427000",
          "content": "<p>You are using sample submission file. Use test set, combined from 4 parquet files instead.</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 722711,
          "author_name": "kaerururu",
          "author_url": "",
          "post_date": "2020-01-19T00:45:59.023000",
          "content": "<p>Oh.. I see. I'll try it, thx!!</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 722991,
          "author_name": "kaerururu",
          "author_url": "",
          "post_date": "2020-01-19T11:21:55.420000",
          "content": "<p>Thank you for your advice!!</p>\n\n<p>Using not feather files but test parquet files resulted in 'reasonable' score. m(_ _)m</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 722245,
      "author_name": "Nicholas Lyu",
      "author_url": "",
      "post_date": "2020-01-18T10:00:14.110000",
      "content": "<p>Same for me!</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 722260,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-01-18T10:24:00.340000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "722351": "Something like that can also happen if you forget to normalize your images in your code. As model is trained on normalized data. So check that also.",
    "722238": "It happened to mee too. In your submission kernel, try to use the train parquet files for a test commit. ",
    "722140": "Maybe there is some problem with your submission kernel??",
    "722375": "During submission the test_images perquet files will be replaced by the complete one. So make your submission for the full test data",
    "722135": "Hi,\n\nAre you deriving test values from test dataset or from sample_submission file, if you predicting values from sample_submission than replace it with data from test parquet files.\n\nAlso it might happen when the algorithm is overfitted than it gives score of .0614.",
    "722132": "I try to make baseline resnet18 model with 100 epoch training, Adam(4e-4), CELoss, no Augmentations.\n\nBut every experiments result in sumple_submission score 0.0614 ..\n\nWhat happens with me !!! Thank you !\n\n",
    "722997": "Thank you for a lot of advices.\n\nIn my case, Not using origianl test parquet but using transformed feather files resluted in this bag. ",
    "722552": "Thank you for many comments !!\nBelow, my latest submission code is here.\n\nsome advices may help me... thank you.\n\nhttps://www.kaggle.com/kaerunantoka/pred-bengali-resnet18-exp5?scriptVersionId=27177286",
    "722245": "Same for me!",
    "722260": ""
  }
}